Providing Post-Hoc Explanation for Node Representation Learning Models Through Inductive Conformal Predictions

Learning with graph-structured data, such as social, biological, and financial networks, requires effective low-dimensional representations to handle their large complex interactions. Recently, the advances of neural networks embedding algorithms, many unsupervised approaches have been proposed for downstream tasks promising results; however, there has limited research on interpreting and, specifically, understanding which parts neighboring nodes contribute representation a node. To mitigate this problem, we propose statistical framework interpret learned representations. Many existing works, are designed supervised node presentation models, compute difference in prediction scores after perturbing edges candidate explanation node; our leverages conformal (CP)-based test verify importance each representation. In evaluation, was verified experimental settings presented results compared those recent baseline methods.